Two-stage atlas subset selection in multi-atlas based image segmentation.

Area of Science:

• Medical image analysis
• Computational anatomy
• Machine learning for medical imaging

Background:

• Large-scale medical image datasets and cloud storage offer opportunities for multi-atlas segmentation.
• Challenges include heterogeneous atlas quality and high computational demands.
• Existing methods struggle with efficiency and accuracy in large-scale scenarios.

Purpose of the Study:

• To develop a novel two-stage method for multi-atlas image segmentation.
• To address challenges of large, heterogeneous atlas collections and computational burden.
• To achieve high segmentation accuracy with reduced computational cost.

Main Methods:

• Proposed a two-stage atlas subset selection scheme.
• Stage 1: Augmented subset selection using low-cost registration and preliminary relevance metric.
• Stage 2: Subset refinement using full registration and refined relevance metric, with an inference model to link metrics.

Main Results:

• The proposed scheme achieved comparable segmentation performance to conventional single-stage methods with significant computation reduction.
• Demonstrated improved Dice similarity coefficients for prostate (0.83, 0.85) and corpus callosum (0.95, 0.95) segmentation compared to alternative methods.
• Results were validated using cross-validation on clinical MRI datasets for prostate and brain segmentation.

Conclusions:

• A novel two-stage atlas subset selection scheme for multi-atlas segmentation was developed.
• The method achieves high segmentation accuracy with substantially reduced computation.
• This approach is well-suited for scenarios involving extensive and heterogeneous atlas collections.